A Threshold-based HMM-DTW Approach for Continuous Sign Language Recognition

Recently, great progress has been made in sign language recognition. Most approaches are based on the Hidden Markov Model (HMM) with various features, such as motion trajectory. Recognition for sign sentences is obtained from optimal path by Viterbi algorithm, however, some wrong jumps are usually caused by transitional movements between signs. To address the problem, in this paper, we propose an approach consisting of two stages: offline training and online recognition. In the offline training stage, we propose a threshold matrix and rate thresholds. Each element of the threshold matrix describes the minimal probability when a segment belongs to a sign, and rate thresholds are defined as the average probability for signs. So, if certain segment's evaluation is smaller than all the thresholds, it is regarded as a transitional movement and then it should be removed. In the online recognition stage, coarse segmentation, based on the threshold matrix, records the time interval for fine segmentation, and fine segmentation, based on Dynamic Time Warping(DTW) and Length-Root method, determines the endpoint for each candidate sign and selects the most possible one. The final recognition is obtained by concatenating the most possible signs. We evaluate our approach with Kinect-based dataset and the experiments demonstrate the effectiveness of our approach.